Memory bandwidth

Memory bandwidth is the rate at which a processor can read data from and write data to its memory, measured in gigabytes or terabytes per second. Peak bandwidth equals the bus width in bits times the data rate per pin, divided by 8 bits per byte. It depends on the memory type, the number of channels or stacks, and the speed grade, not on capacity.

AMD computes it as channels times 8 bytes times the transfer rate. By that formula, an EPYC 9005 socket with 12 channels at 6,400 MT/s peaks at 614 GB/s. An EPYC 9006 with 16 channels at 12,800 MT/s reaches about 1.6 TB/s. The RTX PRO 6000 Blackwell pairs 96 GB of 28 Gbps GDDR7 with a 512-bit bus for 1,792 GB/s. High bandwidth memory (HBM) goes further. The H200 moves 4.8 TB/s from 141 GB of HBM3e, about three times the EPYC 9006 figure.

Rank hardware by GB/s once the working set fits, because any job that rereads its data is paced by bandwidth. One pass over 96 GB takes at least 54 ms at 1,792 GB/s and at least 156 ms at 614 GB/s. Capacity only decides whether the data fits.

Sources

Last verified October 7, 2026.